El Niño and the shifting geography of cholera in AfricaSean M. Moorea,b,c, Andrew S. Azmana, Benjamin F. Zaitchikd, Eric D. Mintze, Joan Brunkarde, Dominique Legrosf,Alexandra Hillf, Heather McKaya, Francisco J. Luquerog,h, David Olsonh, and Justin Lesslera,1

aDepartment of Epidemiology, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD 21205; bDepartment of Biological Sciences, University of NotreDame, Notre Dame, IN 46556; cEck Institute for Global Health, University of Notre Dame, Notre Dame, IN 46556; dDepartment of Earth and Planetary Sciences, JohnsHopkins University, Baltimore, MD 21218; eDivision of Foodborne, Waterborne and Environmental Diseases, Centers for Disease Control and Prevention, Atlanta,GA 30329; fDepartment of Pandemic and Epidemic Diseases, World Health Organization, 1211, Geneva, Switzerland; gEpicentre, 75012 Paris, France; hDepartmentof International Health, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD 21205; and iMédecins Sans Frontières, 75011 Paris, France

Edited by Andrea Rinaldo, Laboratory of Ecohydrology (ECHO/IIE/ENAC), Ecole Polytechnique Federale Lausanne, and approved March 8, 2017 (received forreview October 18, 2016)

The El Niño Southern Oscillation (ENSO) and other climate patternscan have profound impacts on the occurrence of infectious dis-eases ranging from dengue to cholera. In Africa, El Niño conditionsare associated with increased rainfall in East Africa and decreasedrainfall in southern Africa, West Africa, and parts of the Sahel.Because of the key role of water supplies in cholera transmission,a relationship between El Niño events and cholera incidence ishighly plausible, and previous research has shown a link betweenENSO patterns and cholera in Bangladesh. However, there is littlesystematic evidence for this link in Africa. Using high-resolutionmapping techniques, we find that the annual geographic distribu-tion of cholera in Africa from 2000 to 2014 changes dramatically,with the burden shifting to continental East Africa—and awayfrom Madagascar and portions of southern, Central, and WestAfrica—where almost 50,000 additional cases occur during El Niñoyears. Cholera incidence during El Niño years was higher in regionsof East Africa with increased rainfall, but incidence was also higherin some areas with decreased rainfall, suggesting a complex re-lationship between rainfall and cholera incidence. Here, we showclear evidence for a shift in the distribution of cholera incidencethroughout Africa in El Niño years, likely mediated by El Niño’simpact on local climatic factors. Knowledge of this relationshipbetween cholera and climate patterns coupled with ENSO fore-casting could be used to notify countries in Africa when they arelikely to see a major shift in their cholera risk.

cholera | El Niño Southern Oscillation | disease mapping |climate and health | Bayesian mapping

Improvements to water and sanitation have eliminated thethreat of cholera throughout much of the world; however, each

year, millions are infected and over a hundred thousand die inAsia, Africa, and the Caribbean (1). Cholera’s impact may be thegreatest in Africa, where there has been ongoing circulationsince the 1970s (2) and unexpected, explosive epidemics havebeen associated with case fatality rates (CFRs) as high as 50%(commonly 1–15%) (3, 4). Reported CFRs for Africa remaintwice as high as the 2014 global average of 1.2% (5). Choleraepidemics in sub-Saharan Africa have proven difficult to fore-cast, hampering prevention and control efforts (6, 7). However, ithas long been believed that climatic factors in general, and the ElNiño Southern Oscillation (ENSO) in particular, are importantdrivers of cholera incidence (8, 9).ENSO is a periodic, multiannual variation in sea surface

temperatures and winds in the tropical Pacific Ocean that in-fluences weather patterns globally (10, 11). Warm phases in theeastern Pacific Ocean (El Niño events) occur every 2–7 y andare associated with warm sea surface temperatures in parts ofthe western Indian Ocean, above-average rainfall in EastAfrica, and below-average rainfall in dry regions of southernAfrica and the Sahel (Fig. 1A). The 2015–2016 El Niño event isonly the third of the past 40 y to be classified as “strong” or“very-strong” (joining 1982–1983 and 1997–1998; Fig. 1D) (12).The global link between cholera and climate has been thesubject of inquiry since at least the 19th century (13). Most

modern research has focused on South Asia, where interannualvariability in seasonal cholera epidemic size has been associatedwith ENSO strength (8, 14, 15).Warm sea surface temperatures in the Bay of Bengal that often

accompany El Niño events may facilitate the growth of environ-mental reservoirs of Vibrio cholerae, increasing the severity of thatyear’s epidemic (8, 14, 16). In sub-Saharan Africa, however, themajority of cholera cases occur in inland regions (17), hence seasurface temperatures are unlikely to play as direct a role in drivingseasonal and multiannual variations in cholera incidence. There issome evidence that cholera incidence in the Great Lakes Region ofEast Africa increases during abnormally warm El Niño events (18,19). Large cholera epidemics in Africa are also associated with bothvery dry and very wet conditions: dry conditions may force people touse unsafe drinking water sources (17, 20, 21), whereas floodingmay facilitate fecal contamination of drinking water (22). Thiscomplexity combined with the lack of fine-scale data on choleraincidence and environmental covariates has limited our under-standing of how climatic events, like El Niño, impact cholera inci-dence on the continent.Whereas vulnerability to cholera outbreaks is driven by local

conditions, including safe water and sanitation access, healthinfrastructure, and socioeconomic factors, ENSO-related climateperturbations may also modify the distribution of cholera risk.To understand how ENSO affects the geographic distribution ofcholera incidence in Africa, we mapped estimated cholera in-cidence at the scale of 20 × 20 km grid cells throughoutthe continent. Using a hierarchical Bayesian approach, we

Significance

In the wake of the 2015–2016 El Niño, multiple cholera epidemicsoccurred in East Africa, including the largest outbreak since the1997–1998 El Niño in Tanzania, suggesting a link between ElNiño and cholera in Africa. However, little evidence exists forthis link. Using high-resolution mapping techniques, we foundthe cholera burden shifts to East Africa during and following ElNiño events. Throughout Africa, cholera incidence increasedthree-fold in El Niño-sensitive regions, and 177 million peopleexperienced an increase in cholera incidence.Without treatment,the case fatality rate can reach 50%, but accessible, appropriatecare nearly eliminates mortality. Climatic forecasts predicting ElNiño events 6–12 mo in advance could trigger public healthpreparations and save lives.

Author contributions: S.M.M., A.S.A., B.F.Z., E.D.M., F.J.L., and J.L. designed research; S.M.M.,A.S.A., B.F.Z., E.D.M., F.J.L. and J.L. conceived this study; S.M.M., A.S.A., B.F.Z., J.B., D.L.,A.H., H.M., F.J.L., D.O., and J.L. performed research; S.M.M., A.S.A., B.F.Z., D.L., A.H., H.M.,F.J.L., D.O., and J.L. contributed to the collection, assembly, and entry of data; S.M.M., A.S.A.,B.F.Z., and J.L. analyzed data; and S.M.M., A.S.A., and J.L. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.

Freely available online through the PNAS open access option.1To whom correspondence should be addressed. Email: justin@jhu.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1617218114/-/DCSupplemental.

4436–4441 | PNAS | April 25, 2017 | vol. 114 | no. 17 www.pnas.org/cgi/doi/10.1073/pnas.1617218114

integrated data from over 17,000 annual observations of chol-era incidence from 2000 to 2014 in over 3,000 unique locationsof varying spatial extent, ranging from entire countries toneighborhoods. The resulting maps reflect modeled choleraincidence at a fine spatial resolution using reported counts ofcholera cases, key explanatory variables (population density,access to improved drinking water, access to improved sanita-tion, and distance to nearest major water body), and a spa-tially dependent covariance term (Materials and Methods). Wethen examined the potential mechanistic association betweenENSO-related changes in cholera incidence and several envi-ronmental variables including rainfall.

ResultsEl Niño events affect the distribution and magnitude of choleraincidence throughout the African continent (Figs. 1B and 2).Africa as a whole experienced a similar number of cholera casesin El Niño years compared with non-El Niño years between2000–2014 [215,546 (95% credible interval [CrI]: 209,770–

221,704) vs. 209,791 (95% CrI: 202,087–219,047)], but the geo-graphic distribution of these cases fundamentally changed be-tween El Niño and non-El Niño years, with most countriesexperiencing areas of both decreased and increased incidence(Fig. 1B). However, notable regional patterns were observed;southern Africa experienced fewer cholera cases in El Niño years(31,598 fewer cases; 95% CrI: 29,385–33,775), whereas conti-nental East Africa (i.e., excluding Madagascar) had significantincreases in cholera incidence in El Niño years (Fig. 1B), with48,670 (95% CrI: 45,192–52,053) excess cases (SI Appendix,Table S4). Overall, 177 million people live in areas where annualcholera incidence increased by at least 1 per 100,000 during ElNiño years (95% CrI: 166.0–189.5 million), and 81 million live inareas where annual incidence increased by more than 1 per 10,000(95% CrI: 75.8–86.4 million). Likewise, 137 million live in areaswhere annual incidence declined by at least 1 per 100,000 (95%CrI: 125.6–149.4 million) and 69 million live in areas whereannual incidence decreased by more than 1 per 10,000 (95%CrI: 63.2–74.2 million).

1982 1986 1990 1994 1998 2002 2006 2010 2014

−2−1

01

2

3−m

on

th E

NSO

Ind

ex

−60−40−200204060

−10−50510

Rainfalldeviation (%)

Differencelog incidence

20

posi

tive

sens

itivi

tyne

gativ

ese

nsiti

vity

10

0

10

Likelihoodratio

A B C

D

Fig. 1. Geographical distribution of cholera in El Niño and non-El Niño years. (A) Rainfall anomalies in El Niño years, 1980–2015. (B) Fine-scale geographic distributionof cholera anomalies in El Niño years, 2000–2014. (C) Country-level cholera anomalies in El Niño years, based on WHO reports, 1980–2015. The colors represent thelikelihood ratio in support of a significant difference in cholera incidence between El Niño and non-El Niño years. (D) History of strength of ENSO anomalies, 1980–2015. El Niño years are in red, and La Niña years are in blue. Red and blue outlines inA–C represent regions with positive (red) and negative (blue) sensitivity to El Niñoevents in cholera incidence selected by smoothing the normalized difference in cholera incidence using a kernel smoothing algorithmwith a bandwidth of 150 km andthen clustering areas into areas where cholera incidence is positively sensitive, negatively sensitive, and insensitive to El Niño events (Fig. 2).

A

(8) Uganda (western)

(9) Nigeria (Ekiti State)

(7) Democratic Republic of Congo

(east and southeast)

(6) Mozambique

(5) Zambia (Lusaka)

(4) Malawi

(3) Tanzania (TT countrywide)

(2) Kenya (countrywide)

(1) Somolia

B

<0.001

0.01

0.1

1

10

100

1000

10000

Negatively Sensitive Insensitive Positively SensitiveENSO Sensitivity Clusters

Cas

es p

er 1

0,00

0

YearEl Niñonon El Niño

−1.0

−0.5

0.0

0.5

1.0

C Holdout analysiscluster averages

Fig. 2. Geographical distribution and incidence ratesfor El Niño-sensitive regions. (A) Regions with positive(red) and negative (blue) sensitivity to El Niño eventsin cholera incidence. Areas selected by smoothing thenormalized difference in cholera incidence using akernel smoothing algorithm with a bandwidth of150 km, then clustering areas into areas where choleraincidence is positively sensitive (red), negatively sensi-tive (blue) and insensitive (white) to El Niño events.Callouts indicate major reported outbreaks of the2015–16 cholera season (SI Appendix, 1. SI Materialsand Methods). (B) Kernel density (violin) plot of casesper 10,000 in different El Niño-sensitive regions duringEl Niño and non-El Niño years. Black circles are gridcell-level medians ± 1 SD and blue diamonds are grid-cell level means. (C) Overlay of El Niño sensitive clus-ters from holding out each El Niño or non-El Niño pairof years with negatively sensitive clusters = −1 (blue)and positively sensitive clusters = 1 (red).

Moore et al. PNAS | April 25, 2017 | vol. 114 | no. 17 | 4437

ECOLO

GY

STATIST

ICS

To delineate areas that can expect significant increases orreductions in cholera incidence during El Niño years, we classi-fied areas, irrespective of political boundaries, based on thesensitivity of local cholera incidence to El Niño using clusteringand smoothing algorithms. We identified those areas with thelargest increase in expected incidence during El Niño events (thetop quartile) and those with the largest decrease (the bottomquartile, see Materials and Methods). The largest positively sen-sitive cholera cluster extends through continental East Africafrom the horn of Africa down to Mozambique, with smallerclusters scattered throughout the continent (Fig. 2A). The mostdistinct negatively sensitive clusters are in Madagascar andportions of Central and West Africa (Fig. 2A). Overall, 45.8% ofpeople in sub-Saharan Africa live in El Niño-sensitive areas:263.6 million in positively sensitive clusters and 203.7 million innegatively sensitive ones. During El Niño years, cholera inci-dence within positively sensitive clusters increased, on average,almost threefold from 1.1 per 10,000 to 3.3 per 10,000 [relativerate (RR): 2.91; 95% CrI: 2.52–3.19], corresponding to almost55,000 excess cases (Fig. 2B). In negatively sensitive clusters in-cidence decreased from 4.2 per 10,000 to 2.2 per 10,000 (RR:0.53, 95% CrI: 0.48–0.57), a reduction of nearly 40,000 cases. Asensitivity analysis showed that the locations of these clusterswere not driven by a single year or El Niño event, (Fig. 2C).Because we were only able to perform fine-scale mapping of

cholera incidence from 2000 to 2014, there are a limited numberof El Niño events captured in these analyses (2 weak, 2 moder-ate, and 0 strong/very-strong events; SI Appendix, Table S2). Toconfirm that recently observed geographic patterns hold over alonger time scale, we analyzed country-level incidence datareported to the WHO dating back to 1980 (23). Similar to thefine-scale analyses, increases in cholera incidence are concen-trated in continental East Africa, specifically Tanzania andKenya, where El Niño events are associated with increasedrainfall and generally wet conditions (Fig. 1A), whereas thelargest decreases are in Madagascar and Namibia. There is asignificant positive association between ENSO strength andcholera incidence in continental East Africa with above-averagerainfall during El Niño events (Fig. 1 A and C), with incidenceincreasing by 29,226 (95% CrI: 9,403–49,049) cases for every unitincrease in the annual peak ENSO index value (SI Appendix, Fig.S45 and 2. SI Further Results).At the continental scale there was no strong association be-

tween rainfall patterns and cholera incidence. However, areas ofhigher cholera incidence during El Niño years were concentratedin river basins where rainfall anomalies were either below aver-age (RR: 4.4 in basins with rainfall anomalies in the lowestquartile; P < 0.001; 95% CrI: 2.5–7.8) or above average (RR:2.4 in basin with rainfall anomalies in the upper quartile; P =0.003; 95% CrI: 1.4–4.2) compared with areas where rainfallanomalies were insignificant (Fig. 3). In addition, higher choleraincidence during El Niño years was also concentrated in riverbasins with an above-average normalized difference vegetationindex (NDVI), evapotranspiration, and temperature (SI Appen-dix, Figs. S22–S26). The positive association between higherrainfall and higher cholera incidence was concentrated in con-tinental East and southern Africa (encompassing much of thepositively sensitive regions in Fig. 1A), although this associationwas only significant in continental East Africa (RR: 3.5; 95%CrI: 1.4–9.0; Fig. 3C). Cholera incidence in West and CentralAfrica showed no clear positive association with rainfall, butincidence was significantly higher in areas of East Africa (RR:4.7; 95% CrI: 1.6–13.5), West Africa (RR: 6.8; 95% CrI: 2.3–20.3), and Central Africa (RR: 18.7; 95% CrI: 3.7–94.4) withbelow-average rainfall during El Niño years (Fig. 3C).Positive El Niño-associated rainfall anomalies were concen-

trated in positively sensitive cholera clusters. This associationwas due to positive rainfall anomalies in continental East Africa,as rainfall during El Niño years was either normal or below av-erage in the positively sensitive cholera clusters in the othergeographic regions (SI Appendix, Fig. S28). The local association

between cholera incidence and rainfall anomalies was not con-stant across the continent and varied by mean annual rainfalllevels. In drier areas, cholera incidence increased during both lowand high rainfall years, whereas in the wettest areas, cholera in-cidence tends to be below average during high rainfall years (SIAppendix, Fig. S27). The concentration of high-rainfall areas inWest and Central Africa may explain the lack of correlation be-tween increased rainfall and cholera incidence in these regions,whereas the large region of low to moderate rainfall in EastAfrica may be responsible for the positive relationship betweenincreased rainfall and cholera incidence observed during El Niñoevents. Cholera incidence in coastal areas is positively associatedwith sea surface temperature anomalies, particularly whereanomalies are >0.2 °C; however, these regions also have positiverainfall anomalies (SI Appendix, 2. SI Further Results).

DiscussionThe annual reported incidence of cholera in sub-Saharan Africa didnot differ significantly between El Niño and non-El Niño years from2000 to 2014. However, we found evidence of a large-scale shift inincidence within the continent, highlighted by a large increase incholera incidence in East Africa during and after El Niño eventsresulting in almost 50,000 additional annual cases. In addition tothis large positively sensitive region, we identified several other re-gions with increased cholera incidence during El Niño events, aswell as several regions in Central and West Africa that appear to benegatively sensitive to El Niño events, with decreased incidenceduring these years. Although we did not find a strong, continent-wide association between El Niño-associated meteorologicalanomalies and changes in cholera incidence, increased incidencewas associated with positive rainfall anomalies in eastern andsouthern Africa, suggesting that these relationships may vary fromplace to place and should be assessed at a small scale.The 2015–2016 surge in cholera cases in East Africa was not

used in these analyses to define cholera clusters, but, neverthe-less, these outbreaks are concentrated in positively El Niño-sensitive areas (Fig. 2A). Since August 2015, Tanzania hasexperienced a nation-wide outbreak that has infected 25,482 andkilled 299 as of May 17, 2016. Somalia, Kenya, Malawi, Uganda,and Mozambique have also experienced outbreaks since thesummer of 2015 (SI Appendix, 1. SI Materials and Methods).These outbreaks (including Tanzania but excluding Somalia; SIAppendix, 1. SI Materials and Methods) have resulted in at least34,800 reported cases since August 2015, a finding that does notinclude the peak 2016 cholera season in most of these countries.This result equates to nearly 25,000 more cases than would beexpected over the same time period during non-El Niño years,and in-line with the 34,713 cases (95% CrI: 31,043–38,018) es-timated from our analyses for these countries during a moderateEl Niño year. At the edge of the East African positively sensitivecluster, eastern Democratic Republic of Congo has been expe-riencing high cholera incidence, although this trend started be-fore the current El Niño event. Cholera has not, however, beenconfined to positively sensitive clusters. El Niño-insensitive re-gions of western Uganda have recently experienced several smalloutbreaks, as has the Ekiti state of Nigeria and Lusaka, Zambia.The presented shifts in cholera burden associated with El Niño

events are based on estimates of reported incidence, not trueincidence, due to a lack of specific information regarding spatialand temporal reporting biases. Although evidence suggests thereare country-specific reporting biases for cholera that lead to bothoverestimation of cholera incidence during some epidemics andthe underestimation of true cholera incidence in other settings(24), as long as these biases are not temporally variable theyshould not influence the association of cholera with ENSO.However, if the size of outbreaks are systematically under-reported, then the magnitude of the shift in cholera incidenceduring El Niño events may be underestimated due to missedcases. Case over- or underestimation would not shift the di-rection of El Niño-sensitivity unless the quality of surveillanceand reporting efforts were associated with the occurrence of El

4438 | www.pnas.org/cgi/doi/10.1073/pnas.1617218114 Moore et al.

Niño events. We have not found any evidence of increased ordecreased surveillance due to ENSO patterns. However, regionswhere cholera cases are underreported or not reported at all maybe incorrectly categorized as ENSO-insensitive due to a lack ofinformation about when cholera cases occur in these areas.Several recent studies have found that extreme El Niño events

could become more frequent due to climate change (25–27).This finding suggests that shifts in cholera associated with ElNiño events may become more pronounced in the future, per-haps shifting the cholera burden toward East Africa. The be-havior of ENSO under climate change and its future impacts onAfrica are, however, active topics of research, so projections arehighly uncertain (28). Moreover, ENSO is only one way in whichgreenhouse gas induced warming influences the African climate.In the Horn of Africa, for example, El Niño is associated withwet conditions and higher rates of cholera (Fig. 1). However, theHorn of Africa has experienced significant drying in recent de-cades (29). Furthermore, although the majority of global climatemodels project that the region will get wetter in the 21st century,these models have systematic errors in representing the seasonaldistribution of Horn of Africa rainfall and could be producingspurious projections (30). Thus, it is not clear how a potentialincrease in wet El Niño extremes superimposed on a warming-induced drying trend in the Horn of Africa would affect overallcholera risk in that region.Predicting local cholera incidence is a difficult task. However,

here we show clear evidence for a shift in the distribution of choleraincidence throughout regions of Africa in El Niño compared withnon-El Niño years, likely mediated by El Niño’s impact on localclimatic factors. Recent ENSO forecasting models warn of developingEl Niño conditions up to 12 mo in advance (31). As predictive abilityof ENSO anomalies improves (31, 32), our findings provide hopethat we may be able to provide early cholera-risk forecasts. Be-cause effective case management dramatically decreases mortalityin cholera outbreaks, and new control tools (e.g., oral choleravaccines) may prevent, or at least limit, outbreaks, the ability tostep up surveillance, preparedness and response when local risk ishigh can have a significant impact on saving lives. A better under-standing of the mechanisms by which El Niño changes the distri-bution of cholera incidence will help us elucidate the effect ofclimatic change on the global distribution of cholera risk.

Materials and MethodsCholera Data. The cholera data used to generate the fine-scale maps ofcholera incidence were collated from 360 separate datasets (details and dataare available at www.iddynamics.jhsph.edu/projects/cholera-dynamics/data).Annual case counts reported to the WHO from 2000 to 2014 were includedfor each country in sub-Saharan Africa (23). Further details on data sourcesare provided in SI Appendix, 1. SI Materials and Methods. The datasets in-cluded in our analysis included cholera case counts aggregated at varioustime scales from daily to yearly. To estimate annual incidence rates, obser-vations at subannual time scales were aggregated to the annual level, al-though many aggregated annual observations cover only part of a calendaryear. A total of 17,033 annual observations from 3,071 unique locationsfrom 2000 to 2014 were included in the main analysis (Dataset S1). These3,071 unique locations include 44 different countries, 327 first-level admin-istrative units, 1,948 second-level administrative units, and 752 locations atthe third-level administrative unit or lower (SI Appendix, Fig. S1). A summaryof the cholera data used to model the spatial distribution of cholera in-cidence is provided by country (SI Appendix, Tables S1 and S2 and 1. SIMaterials and Methods).

Although our database does not cover 2015 and 2016, we used threeprimary sources for data to provide an overview of the main cholera ep-idemics that have occurred since the onset of the 2015–2016 El Niño event,assumed to be April 1, 2015 (SI Appendix, Table S3). First, we performed aquery on HealthMap.org (March 21, 2016) for the disease cholera using allsources and restricted to Africa. Second, we obtained updated situationreports from the WHO for all available countries. Finally, we used theweekly data reported from UNICEF’s West Africa regional office as of week6, 2016 (33).

−10−5

05

10

Difference inlog incidence

−5

0

5

10

Central Africa East Africa North Africa Southern Africa West Africa

Diff

eren

ce in

log

chol

era

inci

denc

e

Rainfallanomalyquartiles

lowlow−midmid−highhigh

Region

B

C

−20−10

01020

Rainfalldeviation (%)

* * *

*

A

Fig. 3. Association between cholera and rainfall by river basin. (A and B)River basin-level rainfall anomalies (anomalies smaller than ±5% not shown)(A) and (log) cholera anomalies (B) between El Niño and non-El Niño years.(C) River basin-level cholera incidence anomalies by region and the strengthof rainfall anomalies. Positive cholera anomalies are associated with nega-tive rainfall anomalies (lowest quartile) in every region and positive rainfallanomalies (highest quartile) in East and Southern Africa. *The difference inlog cholera incidence between El Niño and non-El Niño years for a givenrainfall anomaly quartile and geographical region is significantly differentfrom the difference in log incidence for areas within that region with rainfallanomalies that fall in the low to mid or mid to high ranges (second and thirdquartiles).

Moore et al. PNAS | April 25, 2017 | vol. 114 | no. 17 | 4439

ECOLO

GY

STATIST

ICS

Covariates and Climate Variables. Gridded population density at a 1km2

resolution for the entire African continent were obtained from WorldPop(www.worldpop.org.uk; accessed March 30, 2016). Distance to the coast wascalculated from a shapefile of the African coastline using the “rgeos”package in R (34). Distances to the nearest large lake or reservoir (surfacearea: >50 km2) and to the nearest permanent smaller water bodies includingrivers (surface area: >0.1 km2) were calculated using the level one and leveltwo layers from the Global Lakes and Wetlands Database (35). The pro-portion of the population with access to improved drinking water (SI Ap-pendix, Fig. S2), improved sanitation (SI Appendix, Fig. S3), and opendefecation were obtained from ref. 36. All covariate data were resampled tothe same 20 km resolution using the “raster” package in R (37).

Gridded rainfall at a spatial resolution of 0.05° from 1981 through 2015 wasobtained from Climate Hazards Group InfraRed Precipitation with Station dataversion 2.0 (38). Rainfall totals were aggregated at an annual time scale run-ning according to the El Niño cycle, which runs from July to June rather thanthe calendar year. Gridded mean temperature, soil moisture, and evapo-transpiration data at a 0.25° spatial resolution were obtained from the GlobalLand Data Assimilation System (39). Annual NDVI values at 0.05° from 2000 to2014 were aggregated from monthly MODIS data (40). Monthly sea surfacetemperatures at a 2.0° spatial resolution from 2000 to 2014 were obtainedfrom the Extended Reconstructed Sea Surface Temperature dataset (41). Foreach climate variable long-term annual means were calculated for the periodsfrom 2000 to 2014 and 1981 to 2014 (rainfall only), and deviations from theselong-term means were calculated for El Niño years. All land-based climatevariables were resampled to 20 × 20 km using the raster package in R (37).

El Niño Analyses. We classified each year as having no El Niño anomaly, or asbeing a weak, moderate or strong El Niño year based on the Oceanic NiñoIndex (ONI) used by the National Oceanic and Atmospheric Administration.Monthly ONI values are calculated from the 3-mo running mean of seasurface temperatures anomalies in the Niño 3.4 region of the Pacific Ocean(5°N–5°S, 120°W–170°W). A year with at least 5 consecutive months with anONI value ≥0.5 °C is classified as an El Niño event; and El Niño events withat least 3 consecutive months ≥1.0 °C, 1.5 °C, or 2.0 °C are classified asmoderate, strong, or very strong events, respectively (Fig. 1D). El Niñoevents typically overlap calendar years, so we classified El Niño years asrunning from July through June of the following calendar year. Becausethe majority of the available cholera data are aggregated at an annuallevel we classified cholera cases occurring during both calendar yearsoverlapping an El Niño event as associated with an El Niño year. For ex-ample, the very strong 1997–1998 El Niño event translates to an 1997 ElNiño year and cholera cases from both 1997 and 1998 were considered tobe associated with the 1997–1998 event.

Using ONI values, from 2000 to 2014, the years 2000, 2001, 2008, 2011,2012, 2013, and 2014 were classified as non-El Niño years. The years from2004 to 2007 were classified as weak El Niño events with ONI valuesof ≥0.5 °C but <1 °C, and 2002–2003 and 2009–2010 were classified asmoderate-to-strong El Niño years with ONI values ≥1 °C for a minimum of3 mo. The analysis presented in the main text included both weak andmoderate-to-strong El Niño events as El Niño years. In SI Appendix, 2.3Sensitivity to the Definition of El Niño Years, we present an analysis usingonly moderate-to-strong El Niño years to understand how the main resultsvary with different El Niño classifications.

Mapping Methodology. A hierarchical Bayesian modeling framework wasused to map aggregated cholera observational data to underlying incidencerates. The entire study region was divided into Nj = 73,979 20 × 20 km gridcells, with any grid cells entirely covered by water or with a populationdensity of 0 excluded from analysis. The annual cholera incidence in eachgrid cell, λj, was modeled using a log-linear regression equation,

log�λj�= β0 + βpXp,j +ψ j ,

with covariates Xp,j. The random effects, ψ j, account for any overdispersionand spatial correlation in the data and are modeled by a conditionalautoregressive distribution (42, 43). Spatial correlation between randomeffects is determined by a binary Nj   X  Nj adjacency matrix, A, with elementaj,k equal to one if grid cells ðj, kÞ are neighbors (sharing an edge), and zerootherwise (and for j= k). The joint distribution of ψ is an Nj-dimensionalmultivariate normal distribution given by

ψ j ∼  N�0, σ2υ ðD− ρAÞ−1

�,

where ρ is a parameter representing the relative strength of spatial

dependence with 0 ≤ ρ < 1 and D is a diagonal matrix with entries

dj,j =PNj

k=1aj,k, where dj,j represents the number of neighbors for grid cell j

(44–46).Each observation, Yi, was mapped to the underlying grid cells that are

within area i and were modeled by a Poisson process:

Yi ∼  PoisðEiÞ.

The expected number of cases, Ei, for each observation is the sum of theexpected number of cases in each grid cell:

Ei = XNi

j=1

λj * pj ,

where pj is the size of the population in grid cell j. Each area-based obser-vation was associated with a polygon and the determination of which gridcells were associated with each observation was performed by using the’extract’ function from the raster library in R (37). For point-based obser-vations, such as geolocated case data or cases from a single refugee camp,the grid cell containing that GPS point was used. Multiple observations forthe same area covering different temporal periods or from different sourceswere treated as independent observations, and data from different, butoverlapping, spatial scales were also treated as independent observations.

The intercept term of the log-linear regression model, β0 and the regressionparameters β = (β1,. . .,βp) were assigned weakly informative Gaussian prior dis-tributions, N(0,10). The spatial autocorrelation parameter ρ was assigned a β (2,1)prior and the precision parameter τv from the spatial autocorrelation term ψ jwasassigned a Γ(0.5,0.0005) prior distribution. The covariates included in our analysiswere level of access to improved drinking water, level of access to improvedsanitation, population density, distance to the nearest coastline, and distance tothe nearest major waterbody. The relationship between cholera incidence andthese covariates was considered separately for El Niño and non-El Niño years todetermine whether their relationship was altered by weather patterns associatedwith the ENSO cycle. Details on model implementation are provided in SI Ap-pendix, and summary model outputs are provided in Datasets S2–S5.

To test the sensitivity of our results to single El Niño or La Niña events, wereran the model while holding out each single pair of years representingeither an El Niño event or a non-El Niño event (eight pairs of years; with theexception of the non-El Niño year 2008, between the 2006–2007 and 2009–2010 El Niño events, where only a single year was withheld from the anal-ysis). The variation in incidence when different El Niño and non-El Niño yearswere withheld are presented in SI Appendix, Figs. S14 and S15. The sensi-tivity of the shift in incidence during El Niño events to holding out differentyears is presented in SI Appendix, Fig. S16.

Clustering of Cholera El Niño-Sensitive Regions. El Niño-sensitive regions wereidentified by smoothing maps of normalized cholera incidence and thenclustering the smoothed incidence by quartiles. Cholera incidence was firstnormalized by dividing the difference in cholera incidence in El Niño versusnon-El Niño years by the square root of the summed variance from El Niñoand non-El Niño years. The distribution function of the normalized choleraincidence was then smoothed using a kernel smoother with a bandwidth of150 km and grid cells weighted by population density (47, 48). Smoothingwas implemented using the image.smooth function in the “fields” R pack-age (49). The results of the smoothing and subsequent clustering with al-ternative bandwidth sizes ranging from 50 to 300 km are shown in SIAppendix, Fig. S17. Grid cells in the lowest quartile were classified as neg-atively sensitive clusters, whereas grid cells in the upper quartile were clas-sified as positively sensitive clusters. Grid cells in the middle quartiles wereclassified as insensitive. The sensitivity of the clustering results to holding outeach pairing of El Niño and non-El Niño years is presented in Fig. 2C.

Country-Level ENSO Anomalies. In addition to using the detailed incidenceestimates data from 2000 to 2014, we also used official country-level annualreports from the WHO dating back to 1980 to understand the relationshipbetween El Niño years and cholera incidence (23). The results presented inFig. 1C are based on a linear regression models for each country of the form:

ffiffiffiffiffiffiCy

q= β0 + β1Iðy ∈ ENÞ+  Ey ,

where, Cy is the number of cases in year y and EN represents the set of yearswith an El Niño event. We explored variants of this model that included alinear term for trends in reporting over time and those using EN sets(i.e., weak and stronger or moderate and stronger). Within Fig. 1C, we

4440 | www.pnas.org/cgi/doi/10.1073/pnas.1617218114 Moore et al.

highlighted countries based on the likelihood ratio comparing models withand without β1, with the darker colors illustrating larger support for a sig-nificant different between El Niño and non-El Niño years.

Association of Cholera with Local Climate. We examined the spatial distri-bution of the following six climatic factors estimated from remote sensingand climate reanalysis in El Niño and non-El Niño years between 2000 and2014: rainfall, temperature, NDVI, soil moisture, evapotranspiration, and seasurface temperature. Of the five land-based factors, deviations in rainfalldiffered by >20% over the largest proportion of the sub-Saharan Africanland area (5.0%; Fig. 3A). Because of its strong association with El Niñoevents, high correlation with other climatic factors, and clear mechanisticrelationship with cholera transmission, here we focus primarily on the re-lationship between rainfall and cholera (analyses of other climatic factorsare included in SI Appendix, 2. SI Further Results). The association betweenrainfall and cholera was examined by aggregating ENSO-associated rainfallanomalies and incidence at the river-basin scale because of the impact ofrainfall within a river basin on local surface water and flooding. The majorriver basins of Africa, along with their subbasins used in this analysis, wereobtained from the World Wildlife Fund HydroSHEDS project (50). Becausewe did not observe a simple linear relationship between ENSO-associatedshifts in climate and cholera at the continental- or basin-scale, the relationshipbetween climate measures and shifts in cholera incidence at the basin-scale

were tested by aggregating climate anomalies into quartiles and then com-paring shifts in cholera incidence in the areas with climate anomalies in thelower and upper quartiles to areas not experiencing significant anomalies(middle quartiles) using one-way ANOVA tests. These statistical tests allow usto determine whether large ENSO-associated negative or positive shifts inclimate variables such as rainfall are associated with shifts in cholera incidence.

ACKNOWLEDGMENTS. We thank Abdinasar Abubakar for providing in-formation about current cholera outbreaks; Carla Zelaya, Marisa Hast, KerryShannon, Susan Fallon, Chris Troeger, and Yuru Huang for assistance withidentifying, extracting, and entering data; and the Johns Hopkins InfectiousDisease Dynamics group for feedback on study design and analysis. Researchby S.M.M., A.S.A., H.M., and J.L. was supported by Bill and Melinda GatesFoundation Grant OPP1127318. B.F.Z. was supported by National ScienceFoundation Grant 1639214 [“Innovations at the Nexus of Food, Energy, andWater Systems (Track 1): Understanding multi-scale resilience options forvulnerable regions”]. Cholera data were provided by the Ministries of Healthof Benin, Democratic Republic of Congo, Mozambique, South Sudan, andNigeria. Additional cholera data were provided by Médecins Sans Frontières(MSF) and MSF/Epicentre, the WHO, and the United Nations Relief Agency.Data are indexed at www.iddynamics.jhsph.edu/projects/cholera-dynamics/data and are either directly available or available by request of the owningagency.

1. World Health Organization (2015) Cholera, 2014. Wkly Epidemiol Rec 90:517–528.2. Mutreja A, et al. (2011) Evidence for several waves of global transmission in the

seventh cholera pandemic. Nature 477:462–465.3. Siddique AK, et al. (1995) Why treatment centres failed to prevent cholera deaths

among Rwandan refugees in Goma, Zaire. Lancet 345:359–361.4. Mengel MA, Delrieu I, Heyerdahl L, Gessner BD (2014) Cholera outbreaks in Africa.

Curr Top Microbiol Immunol 379:117–144.5. WHO (2015) Cholera case fatality rate: Situations and trends. Available at www.who.

int/gho/epidemic_diseases/cholera/case_fatality_rate_text/en. Accessed May 5, 2016.6. Constantin de Magny G, Guégan J-F, Petit M, Cazelles B (2007) Regional-scale climate-

variability synchrony of cholera epidemics in West Africa. BMC Infect Dis 7:20.7. Reyburn R, et al. (2011) Climate variability and the outbreaks of cholera in Zanzibar,

East Africa: A time series analysis. Am J Trop Med Hyg 84:862–869.8. Pascual M, Rodó X, Ellner SP, Colwell R, Bouma MJ (2000) Cholera dynamics and El

Niño-Southern Oscillation. Science 289:1766–1769.9. Patz JA, Campbell-Lendrum D, Holloway T, Foley JA (2005) Impact of regional climate

change on human health. Nature 438:310–317.10. Rasmusson EM, Wallace JM (1983) Meteorological aspects of the el nino/southern

oscillation. Science 222:1195–1202.11. Ropelewski CF, Halpert MS (1987) Global and Regional Scale Precipitation Patterns

Associated with the El Niño/Southern Oscillation. Mon Weather Rev 115:1606–1626.12. National Oceanic and Atmospheric Administration (2016) Climate Prediction Center: ENSO

diagnostic discussion. Available at www.cpc.ncep.noaa.gov/products/analysis_monitoring/enso_advisory/ensodisc.html. Accessed April 6, 2016.

13. Bellew HW (1885) The History of Cholera in India from 1862-1881: Being a Descriptiveand Statistical Account of the Disease as Derived from the Published Official Reportsof the Several Provincial Governments During that Period and Mainly in Illustration ofthe Relation Between Cholera Activity and Climatic Conditions Together with Origi-nal Observations on the Causes and Nature of Cholera (Trubner & Co., London).

14. Colwell RR (1996) Global climate and infectious disease: The cholera paradigm.Science 274:2025–2031.

15. Colwell RR, Huq A (1998) Global microbial ecology: Biogeography and diversity ofVibrios as a model. J Appl Microbiol 85:134S–137S.

16. Koelle K, Rodó X, Pascual M, Yunus M, Mostafa G (2005) Refractory periods and cli-mate forcing in cholera dynamics. Nature 436:696–700.

17. Rebaudet S, Sudre B, Faucher B, Piarroux R (2013) Environmental determinants ofcholera outbreaks in inland Africa: A systematic review of main transmission foci andpropagation routes. J Infect Dis 208:S46–S54.

18. Bompangue Nkoko D, et al. (2011) Dynamics of cholera outbreaks in Great Lakesregion of Africa, 1978-2008. Emerg Infect Dis 17:2026–2034.

19. Olago D, et al. (2007) Climatic, socio-economic, and health factors affecting humanvulnerability to cholera in the Lake Victoria basin, East Africa. Ambio 36:350–358.

20. Umoh JU, Adesiyun AA, Adekeye JO, Nadarajah M (1983) Epidemiological features ofan outbreak of gastroenteritis/cholera in Katsina, Northern Nigeria. J Hyg (Lond) 91:101–111.

21. Tauxe RV, Holmberg SD, Dodin A, Wells JV, Blake PA (1988) Epidemic cholera in Mali:High mortality and multiple routes of transmission in a famine area. Epidemiol Infect100:279–289.

22. Griffith DC, Kelly-Hope LA, Miller MA (2006) Review of reported cholera outbreaksworldwide, 1995-2005. Am J Trop Med Hyg 75:973–977.

23. World Health Organization (2014) Weekly epidemiological record: Cholera articles.Available at www.who.int/cholera/statistics/en. Accessed March 30, 2016.

24. Ali M, Nelson AR, Lopez AL, Sack DA (2015) Updated global burden of cholera inendemic countries. PLoS Negl Trop Dis 9:e0003832.

25. Cai W, et al. (2014) Increasing frequency of extreme El Niño events due to greenhousewarming. Nat Clim Chang 4:111–116.

26. Latif M, Semenov VA, ParkW (2015) Super El Niños in response to global warming in aclimate model. Clim Change 132:489–500.

27. Santoso A, et al. (2013) Late-twentieth-century emergence of the El Niño propagationasymmetry and future projections. Nature 504:126–130.

28. Wang C, Deser C, Yu J-Y, DiNezio P, Clement A (2017) Coral reefs of the easterntropical Pacific. El Niño and Southern Oscillation (ENSO): A Review. Coral Reefs of theEastern Tropical Pacific, eds Glynn PW, Manzello DP, Enochs IC (Springer, Dordrecht,The Netherlands), pp 85–106.

29. Williams AP, et al. (2011) Recent summer precipitation trends in the Greater Horn ofAfrica and the emerging role of Indian Ocean sea surface temperature. Clim Dyn 39:2307–2328.

30. Tierney JE, Ummenhofer CC, deMenocal PB (2015) Past and future rainfall in the Hornof Africa. Sci Adv 1:e1500682.

31. Ludescher J, et al. (2014) Very early warning of next El Niño. Proc Natl Acad Sci USA111:2064–2066.

32. Barnston AG, Tippett MK, L’Heureux ML, Shuhua L, DeWitt DG (2012) Skill of real-time seasonal ENSO model predictions during 2002–11: Is our capability increasing?Bull Am Meteorol Soc 93:631–651.

33. United Nations International Children’s Emergency Fund (2016) Cholera outbreaks inCentral and West Africa: 2016 regional update - Week 06. Available at https://www.unicef.org/cholera/files/WCA_Cholera_Update_W6.pdf. Accessed April 6, 2016.

34. Bivand R, Rundel C (2014) rgeos: Interface to Geometry Engine - Open Source (GEOS).Available at cran.r-project.org/web/packages/rgeos/index.html Accessed November16, 2016.

35. Lehner B, Bernhard L, Petra D (2004) Development and validation of a global data-base of lakes, reservoirs and wetlands. J Hydrol (Amst) 296:1–22.

36. Pullan RL, Freeman MC, Gething PW, Brooker SJ (2014) Geographical inequalities inuse of improved drinking water supply and sanitation across sub-Saharan Africa:Mapping and spatial analysis of cross-sectional survey data. PLoS Med 11:e1001626.

37. Hijmans RJ (2015) raster: Geographic Data Analysis and Modeling. Available at https://cran.r-project.org/web/packages/raster/index.html. Accessed November 16, 2016.

38. Funk C, et al. (2015) The climate hazards infrared precipitation with stations–A newenvironmental record for monitoring extremes. Sci Data 2:150066.

39. Rodell M, et al. (2004) The Global Land Data Assimilation System. Bull Am MeteorolSoc 85:381–394.

40. Didan K (2015) MOD13C2 MODIS/Terra Vegetation Indices Monthly L3 Global0.05Deg CMG V006, https://dx.doi.org/10.5067/MODIS/MOD13C2.006.

41. Smith TM, Reynolds RW, Peterson TC, Jay L (2008) Improvements to NOAA’s historicalmerged land–ocean surface temperature analysis (1880–2006). J Clim 21:2283–2296.

42. Besag J, Julian B, Jeremy Y, Annie M (1991) Bayesian image restoration, with twoapplications in spatial statistics. Ann Inst Stat Math 43:1–20.

43. Banerjee S, Carlin BP, Gelfand AE (2014) Hierarchical Modeling and Analysis forSpatial Data (CRC Press, Boca Raton, FL), 2nd Ed.

44. Stern HS, Cressie N (2000) Posterior predictive model checks for disease mappingmodels. Stat Med 19:2377–2397.

45. White G, Ghosh SK (2009) A stochastic neighborhood conditional autoregressivemodel for spatial data. Comput Stat Data Anal 53:3033–3046.

46. Lee D (2011) A comparison of conditional autoregressive models used in Bayesiandisease mapping. Spat Spatio-Temporal Epidemiol 2:79–89.

47. Nadaraya EA (1964) On estimating regression. Theory Probab Appl 9:141–142.48. Watson GS (1964) Smooth regression analysis. Sankhya Ser A 26:359–372.49. Nychka D, Furrer R, Sain S (2015) fields: Tools for Spatial Data. Available at https://

cran.r-project.org/web/packages/fields/index.html. Accessed November 16, 2016.50. Lehner B, Bernhard L, Günther G (2013) Global river hydrography and network

routing: Baseline data and new approaches to study the world’s large river systems.Hydrol Processes 27:2171–2186.

Moore et al. PNAS | April 25, 2017 | vol. 114 | no. 17 | 4441

ECOLO

GY

STATIST

ICS